Explainable, data-driven monitoring of scrubber washwater chemistry: a compliance framework for exhaust gas cleaning systems in chemical engineering applications

Exhaust gas cleaning systems are now routinely installed on ships to comply with international sulfur emission limits. While their performance on air emissions is well understood, monitoring of scrubber washwater discharges is still limited in practice. In many cases, operational measurements are available, but their connection to regulatory compliance remains indirect or unclear. In the proposed approach, we focus on how existing shipboard data can be used more effectively for washwater compliance assessment. We propose a data-driven and explainable approach that combines physics-based proxy calculations with standard compliance rules, anomaly detection, and machine-learning tools. High-frequency operational data from several diesel generator units are used to estimate washwater acidity when continuous pH measurements are not available. The proxy calculation relies on engine load, washwater flow rate, and seawater buffering capacity. This makes it possible to follow discharge conditions continuously and compare them with both international and United States regulatory limits. The analysis shows a clear difference between steady and transitional operation. During stable operation at moderate to high engine loads, typically between 70 and 75 percent of rated capacity, washwater chemistry remains stable. In these periods, pH values stay within a narrow range, generally between 5.8 and 6.5. Short departures from compliance are observed mainly during transient phases, such as startup, shutdown, or rapid load changes. These events are brief and usually last only a few minutes. An anomaly detection method based on interquartile ranges is able to identify these excursions without reacting to isolated sensor noise. On top of this analytical layer, we introduce a supervised classification model as an early-warning tool. The model is trained using a set of temporal and operational features derived from the data. Under cross-validation, it achieves a mean area under the receiver operating characteristic curve of 0.89 and a recall of 98 percent. An explainable analysis of the model output shows that predicted violations are primarily associated with changes in buffering capacity, variations in washwater flow, and short-term pH behavior. These results suggest that washwater compliance risks are not random. They are closely linked to operating regimes and to specific system transitions. The proposed approach illustrates how physics-informed and explainable analytics can turn routine shipboard data into practical and auditable compliance information, useful both for ship operators and for regulatory oversight.